Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Multimodal PDE Foundation Model for Prediction and Scientific Text Descriptions

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A multimodal transformer trained on 52 parametric ODE/PDE families predicts solutions with 3.21% average in-distribution relative error (7.78% for 20%-range out-of-distribution parameters) while generating scientifically consistent text…

desk verdict A useful proof-of-concept that adds text generation to a multimodal operator learner, but the abstract's 'incomplete symbolic input' claim is untested and should be trimmed. read the letter →

arxiv 2502.06026 v1 pith:HKGNLZZC submitted 2025-02-09 cs.LG cs.NAmath.NA

Elisa Negrini , Yuxuan Liu , Liu Yang , Stanley J. Osher , Hayden Schaeffer This is my paper · ORCID
classification cs.LGcs.NAmath.NA MSC 68T0768T5065M99
keywords multimodalfoundationmodelPDEoperatorlearningtransformerODE/PDEpredictionscientifictextgenerationconservationlawsout-of-distributiongeneralizationGPT-2fine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that one transformer-based model can act as a general-purpose solver for many differential equations at once, and can do so while also producing readable scientific text about the equation or its solution. The model is trained on 52 parametric ODE and PDE families, taking in text that states the governing equation plus numerical values for parameters and initial conditions, and returning both predicted solution values at arbitrary query points and natural-language descriptions. On in-distribution test data the average relative prediction error is 3.21%, and it stays below 7.8% when test parameters are sampled from a 20%-wide range around the training values. The text output reaches a semantic-similarity F1 score of 0.937 on average, and a manual check of 100 generated descriptions found them consistent with the input every time. The demonstrations use prompts in which the equation is explicitly written as a formula. If the claim holds, it points toward foundation models that are simultaneously numeric solvers and interpretable scientific assistants.

What carries the argument

The load-bearing object is a multimodal token sequence processed by a fine-tuned generative language model backbone. Text tokens come from a standard language-model tokenizer, while numerical data (initial condition and parameters) are encoded by a small multilayer perceptron into continuous feature vectors; the mixed sequence is then fed through the language model. Numerical solutions are produced by a separate cross-attention data decoder that treats query locations as independent evaluation points, in the spirit of branch-and-trunk operator networks, so cost scales linearly with the number of query points. Text is produced autoregressively from the same backbone by greedy next-token prediction, and the total loss is a weighted sum of a relative squared error for the numerical output and a cross-entropy loss for the text.

What would settle it

Hand the trained model a natural-language description that does not contain the formula—for instance, 'a wave that spreads and smooths over time'—and compare its numerical prediction to the true heat-equation solution. If accuracy collapses to the level of a model that ignored the text, the multimodal fusion is not doing the claimed work; if it stays accurate, the claim generalizes beyond formula-style prompts.

Watch

Extended reading notes

Core claim

The paper's central claim is that numerical and textual modalities can be fused in a single transformer so that operator learning and text generation reinforce each other. Given a prompt such as 'The given equation is $u_t = c u_{xx}$ where $c = 0.003$ and $u(x,0) = u_0(x)$,' the model outputs the solution at user-specified query locations and a sentence describing, for example, that the heat equation is a parabolic PDE modeling heat spread. The authors report that this works across linear and nonlinear ODEs, reaction-diffusion systems, conservation laws, and other PDE families, and that the model can also identify whether a conservation-law solution will develop shocks or rarefactions. They further show that the trained model extrapolates in time for several equation classes by feeding its own predicted solution at $t=5$ as the new initial condition, with relative errors mostly under 13%.

Load-bearing premise

The load-bearing premise is that the text prompt uniquely identifies the governing equation: every experiment uses an input in which the equation is explicitly written as a formula, so if real-world text descriptions are vague or incomplete, the model has not been shown to know which equation to solve.

Editorial extensions

If this is right

  • A single trained model can approximate solutions for dozens of equations at once, instead of one network per equation.
  • Because the text prompt names the equation, the numerical input can be minimal—just the initial condition and parameters—and the model still generalizes to new parameter values within and beyond the training range.
  • The cross-attention data decoder allows solution evaluation at arbitrary, independent query points, so spatial and temporal resolution can be chosen after training.
  • The text channel can carry scientifically useful classifications, such as whether a conservation-law solution forms a shock or a rarefaction, alongside the numerical solution.
  • Time extrapolation is possible for some equation classes by iterating the model with its own output as the next initial condition, though sharp-transition problems such as advection remain much harder.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the approach becomes most distinctive when the text modality does real work—for example, identifying the equation or solution regime from a verbal description alone; the paper leaves that regime untested, since all prompts explicitly state the equation.
  • Editorial inference: one could make a falsifiable test of the fusion mechanism by training the same architecture with text prompts replaced by random labels; if numerical accuracy does not drop, the text channel is not carrying the equation information.
  • Editorial inference: the 100%-consistent text results were verified on 100 examples only; a larger human evaluation across all 52 families would be the natural next check before relying on the descriptions.
  • Editorial inference: the schema suggests a direct extension—conditioning text generation on a user-specified aspect such as 'describe the long-time behavior'—which the authors themselves note as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a multimodal transformer-based foundation model for ODEs/PDEs. The architecture combines a pretrained GPT-2 backbone, an MLP-based encoder for numerical parameters and initial conditions, and a cross-attention data decoder for operator-style numerical predictions, together with autoregressive text generation. The model is trained on a synthetic dataset of 52 parametric ODE/PDE families with GPT-4-generated scientific text descriptions. The reported results include 3.21% average relative numerical error on in-distribution test data, 7.78% on 20%-range out-of-distribution data, a BERTScore-F1 of 0.937 for generated text, and temporal extrapolation on a selected subset of equations.

Significance. If the central claim is fully supported, the paper would make a useful contribution to multimodal PDE foundation models by adding free-text input and output alongside numerical solution prediction. The paper's strengths are its relatively broad synthetic benchmark (52 equation families), the inclusion of code, and the useful comparison context with FNO/DeepONet in Section 4.6. However, the paper's main advertised advantage—handling settings where symbolic representations may be incomplete or unavailable—is not evaluated at all: every experiment supplies the governing equation as an explicit formula in the text prompt. The numerical results are therefore consistent with the model performing equation-family recognition from the explicit formula, and the claimed multimodal advantage over symbolic-input methods such as PROSE is unsubstantiated. The text-generation evaluation is also weaker than the abstract suggests, relying on semantic similarity against a random reference description rather than a correctness measure.

major comments (4)
  1. [Section 1 and Sections 4.3-4.6] The abstract and introduction claim the model can handle settings where symbolic representations may be incomplete or unavailable, but no experiment tests this regime. The paper explicitly states in Section 1: 'In our experiments, we primarily used the first type of input, where the equation is explicitly provided as a formula.' All numerical experiments in Sections 4.3, 4.5, and 4.6, and all text examples in Table 5, include the exact governing equation in the input prompt (e.g., 'The equation is u_t + q^2 u_xxx + u u_x = 0'). There is no experiment with a descriptive prompt such as 'solve the heat equation with coefficient 0.003', nor with incomplete, paraphrased, or noisy descriptions. Because the model always receives the symbolic equation as text, the numerical accuracy and text outputs could be driven entirely by matching the explicit formula to a known family, so the claimed advantage over PROSE, which already takes symbolic expressions as input, is not established. I recommend adding experiments with descriptive or partial prompts (including, if feasible, training on such prompts) or clearly re-scoping the claims to the explicit-formula setting.
  2. [Section 4.6 and Table 6] The temporal extrapolation claim is supported only on a favorable subset of equations, and the selection rule is not stated. The text says results are shown 'for certain classes of equations where the solution at time t = 5 is similar to the initial condition used for training', but the reader is not told how this similarity was measured or which classes were excluded and why. Without this information, the extrapolation results in Table 6 are not a well-defined test of the model's generalization capability. Additionally, no error bars or multiple-run statistics are reported anywhere in the paper, so it is impossible to tell whether differences between classes (e.g., 7.94% vs. 29.3%) are meaningful or noise.
  3. [Section 4.4 and Table 3] The text-generation evaluation is not a correctness evaluation. BERTScore compares the generated sentence to a randomly selected reference description from a set of 50, so a semantically different but correct description can receive a low score, while a superficially similar but factually wrong description could receive a high score. The manual verification of 100 test examples is useful, but the protocol is not described: the selection of examples, the criteria for judging consistency, and whether the evaluators were blinded or independent are all absent. Given the paper's claim that correct descriptions were generated '100% of times', this evaluation needs to be made systematic, or the claim should be softened to '100% of the 100 manually checked examples'.
  4. [Section 4.6 'Comparison with Standard Operator Learning'] The comparison with FNO and DeepONet is not a controlled baseline for the multimodal claim. The cited numbers come from previous papers [25, 39] where FNO/DeepONet receive multiple input snapshots but not the equation text, whereas the proposed model receives the explicit equation in the prompt. This comparison confounds the input modality with the architecture and training data, so it does not show that the textual modality is responsible for the improved accuracy. A more informative baseline would be the same architecture with the equation text removed or replaced by a generic placeholder, or with the equation given in a symbolic modality.
minor comments (5)
  1. [Section 4.6, Table 6] The sentence 'the extrapolation error is < 13% for most equation classes' is imprecise: five classes are reported and the Advection Equation error is 29.3%, so the statement should say 'for four of the five tested classes' or otherwise account for the advection exception.
  2. [Appendix A] The dataset details are incomplete: Table 1 lists conservation-law variants with indices 35-52 (e.g., 'Burgers-Inviscid Conservation law Cosine Flux with one shock'), but Appendix A does not give the equations or generation settings for these indices. Please add the missing definitions.
  3. [Section 4.2] The reported relative error is defined as ||u_hat - u||/||u||, while the training loss in Section 3.5 is a squared relative error. The relationship between the two should be clarified, and it should be stated explicitly whether the reported errors are computed over the full space-time grid or only at the query points.
  4. [Section 3.1 and Section 4.1] Several experimental details needed for reproducibility are missing: the number of training/validation/test splits, the values of the loss weights alpha and beta, learning rate, batch size, number of epochs, and the exact tokenization/encoding scheme for numerical input. The GitHub link is a good start, but the paper should report these core hyperparameters.
  5. [Figures 2 and 3] The figures are dense and the axis labels/color bars are not legible at the current resolution. In particular, the 'absolute difference' panels would benefit from a shared color scale so that errors are visually comparable across rows.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the numerical and text-generation claims are empirical supervised tasks evaluated against independent ground-truth solvers and external references, not derivationally forced by their inputs.

full rationale

This is an empirical machine-learning study rather than a derivation chain, and no step exhibits a stated output that is equivalent by construction to an input or to a fitted parameter renamed as a prediction. Numerical solution errors are computed against ground-truth solutions generated by classical numerical solvers (SciPy for ODEs and the methodology of [39] for PDEs), so the reported 3.21% in-distribution and 7.78% 20%-OOD errors are genuine held-out evaluations and do not reduce to the training inputs. The text-generation results are likewise a supervised next-token prediction task trained with cross-entropy against GPT-4-generated descriptions and evaluated with BERTScore plus a 100-example manual consistency check; the use of GPT-4 to produce training labels is a data-provenance and validity concern, not a circular reduction, because the model outputs are not used to define their own loss or correctness criterion. The paper does cite prior work by the same group, notably [25] and [39], for dataset-generation methodology and baseline FNO/DeepONet comparisons, but these citations are not load-bearing for the central claim: the multimodal model's errors would stand or fall on the reported solver-computed ground truths independent of those citations. Finally, the admitted gap that all experiments use explicit formula text inputs (Section 1: 'In our experiments, we primarily used the first type of input, where the equation is explicitly provided as a formula') is an external-validity or overclaim issue about generalization to vague descriptions, not a circularity in the claims that are actually tested. No quoted reduction between input and output exists, so the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and does not rely on a mathematical derivation, so the ledger contains design choices and domain assumptions rather than fitted constants. The free parameters are hyperparameters and dataset construction choices that affect the numbers but are not derived from first principles. The key domain assumptions are periodic boundary conditions, the reliability of GPT-4 as text label source, the accuracy of external numerical solvers, and the completeness of the text prompt.

free parameters (4)
  • Loss weight alpha (numerical) = not reported
    Hyperparameter balancing relative L2 loss for the operator output; its value affects trade-off between numerical and text losses, but the paper does not state it.
  • Loss weight beta (text) = not reported
    Hyperparameter balancing cross-entropy text loss; same as above.
  • Number of spatial discretization points for PDEs = 128
    The paper discretizes the spatial domain into 128 points (Section 4.2). This fixed grid is a design choice that affects the numerical error and not a derived quantity.
  • Time interval for training = [0,5]
    All training data is generated on t in [0,5]; extrapolation is tested beyond this interval. The choice is arbitrary and part of the experimental setup.
assumptions (4)
  • domain assumption All experiments assume periodic boundary conditions
    Section 1 states 'We will assume periodic boundary conditions for uniformity in experiments.' The model is never tested on other boundary conditions.
  • domain assumption GPT-4-generated text descriptions are treated as correct ground truth
    Section 4.4: 'the text descriptions for training were generated using GPT-4'. The model's text output is judged against these AI-generated references, so correctness is relative to GPT-4's prose.
  • domain assumption Numerical ground truth from SciPy and the solver of [39] is accurate
    Appendix A: ODE solutions use SciPy's solve_ivp; PDEs follow the methodology of [39]. If solvers are inaccurate or under-resolved, the reported errors are biased.
  • domain assumption The text prompt fully specifies the equation
    All experiments provide the equation as an explicit formula in the text prompt (Section 4.3). The model is not tested on vague descriptions, so the multimodal value depends on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Multimodal PDE Foundation Model for Prediction and Scientific Text Descriptions." pith.science (2026). https://pith.science/paper/HKGNLZZC

@misc{pith2026250206026,
  author       = {Pith},
  title        = {Pith review of: A Multimodal PDE Foundation Model for Prediction and Scientific Text Descriptions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HKGNLZZC}},
  note         = {Machine review of arXiv:2502.06026}
}
read the original abstract

Neural networks are one tool for approximating non-linear differential equations used in scientific computing tasks such as surrogate modeling, real-time predictions, and optimal control. PDE foundation models utilize neural networks to train approximations to multiple differential equations simultaneously and are thus a general purpose solver that can be adapted to downstream tasks. Current PDE foundation models focus on either learning general solution operators and/or the governing system of equations, and thus only handle numerical or symbolic modalities. However, real-world applications may require more flexible data modalities, e.g. text analysis or descriptive outputs. To address this gap, we propose a novel multimodal deep learning approach that leverages a transformer-based architecture to approximate solution operators for a wide variety of ODEs and PDEs. Our method integrates numerical inputs, such as equation parameters and initial conditions, with text descriptions of physical processes or system dynamics. This enables our model to handle settings where symbolic representations may be incomplete or unavailable. In addition to providing accurate numerical predictions, our approach generates interpretable scientific text descriptions, offering deeper insights into the underlying dynamics and solution properties. The numerical experiments show that our model provides accurate solutions for in-distribution data (with average relative error less than 3.3%) and out-of-distribution data (average relative error less than 7.8%) together with precise text descriptions (with correct descriptions generated 100% of times). In certain tests, the model is also shown to be capable of extrapolating solutions in time.

Figures

Figures reproduced from arXiv: 2502.06026 by the authors.

Figure 1
Figure 1. Model Illustration. Our model processes multimodal input, where textual prompts describe equations, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Comparison of Outputs: For PDE examples, we show from left to right the ground truth solution, the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Example of extrapolation in time. The model is trained on data generated for [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. PDEformer-2: A Versatile Foundation Model for Two-Dimensional Partial Differential Equations

    math.NA 2025-07 conditional novelty 6.0 of 10

    PDEformer-2 is a pretrained graph-transformer and implicit-neural-representation model that solves a broad class of 2D PDEs from their symbolic form, with zero-shot, few-shot, and inverse-problem capabilities.

Reference graph

Works this paper leans on

51 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Neural machine translation by jointly learning to align and translate

    Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014

    arXiv 2014

  3. [3]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, and others. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    work page 1901

  4. [4]

    Choose a transformer: Fourier or galerkin

    Shuhao Cao. Choose a transformer: Fourier or galerkin. Advances in neural information processing systems , 34:24924–24940, 2021

    work page 2021

  5. [5]

    Vicon: Vision in-context operator networks for multi-physics fluid dynamics prediction

    Yadi Cao, Yuxuan Liu, Liu Yang, Rose Yu, Hayden Schaeffer, and Stanley Osher. Vicon: Vision in-context operator networks for multi-physics fluid dynamics prediction. arXiv preprint arXiv:2411.16063 , 2024

    arXiv 2024

  6. [6]

    Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems

    Tianping Chen and Hong Chen. Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems. IEEE transactions on neural networks , 6(4):911– 917, 1995

    work page 1995

  7. [7]

    W. Chen, J. Song, P. Ren, S. Subramanian, D. Morozov, and M. W. Mahoney. Data-efficient operator learning via unsupervised pretraining and in-context learning. arXiv preprint arXiv:2402.15734 , 2024

    work page Pith review arXiv 2024

  8. [8]

    Machine learning for numerical weather and climate modelling: a review

    Catherine O de Burgh-Day and Tennessee Leeuwenburg. Machine learning for numerical weather and climate modelling: a review. Geoscientific Model Development , 16(22):6433–6477, 2023

    work page 2023

Show all 51 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  2. [10]

    Palm-e: An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023

    2023 arXiv

  3. [11]

    Deep multi-modal object detection and semantic segmentation for au- tonomous driving: Datasets, methods, and challenges

    Di Feng, Christian Haase-Sch¨ utz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for au- tonomous driving: Datasets, methods, and challenges. IEEE Transactions on...

    2020

  4. [12]

    Foundation models in robotics: Applications, challenges, and the future

    Roya Firoozi, Johnathan Tucker, Stephen Tian, Anirudha Majumdar, Jiankai Sun, Weiyu Liu, Yuke Zhu, Shuran Song, Ashish Kapoor, Karol Hausman, et al. Foundation models in robotics: Applications, challenges, and the future. The International Journal of Robotics Research , page 0...

    2023

  5. [13]

    xval: A continuous number encoding for large language models

    Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, et al. xval: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989 , 2023

    arXiv

  6. [14]

    Pre-trained models: Past, present and future

    Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu, Yuan Yao, Ao Zhang, Liang Zhang, et al. Pre-trained models: Past, present and future. AI Open, 2:225–250, 2021

    2021

  7. [15]

    Herde, B

    M. Herde, B. Raoni´ c, T. Rohner, R. K¨ appeli, R. Molinaro, E. B´ ezenac, and S. Mishra. Poseidon: Efficient foundation models for pdes. arXiv preprint arXiv:2405.19101 , 2024

    2024 arXiv

  8. [16]

    A comprehensive survey of regression-based loss functions for time series forecasting

    Aryan Jadon, Avinash Patil, and Shruti Jadon. A comprehensive survey of regression-based loss functions for time series forecasting. In International Conference on Data Management, Analytics & Innovation , pages 117–147. Springer, 2024. 14

    2024

  9. [17]

    Jollie, J

    D. Jollie, J. Sun, Z. Zhang, and H. Schaeffer. Time-series forecasting, knowledge distillation, and refinement within a multimodal pde foundation model. arXiv preprint arXiv:2409.11609 , 2024

    2024 arXiv

  10. [18]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, volume 1, page 2. Minneapolis, Min- nesota, 2019

    2019

  11. [19]

    Importance of search and evaluation strategies in neural dialogue modeling

    Ilia Kulikov, Alexander H Miller, Kyunghyun Cho, and Jason Weston. Importance of search and evaluation strategies in neural dialogue modeling. arXiv preprint arXiv:1811.00907 , 2018

    2018 arXiv

  12. [20]

    Visualbert: A simple and performant baseline for vision and language

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 , 2019

    1908 arXiv

  13. [21]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  14. [22]

    Accelerated replica exchange stochastic gradient langevin diffusion enhanced bayesian deeponet for solving noisy parametric pdes

    Guang Lin, Christian Moya, and Zecheng Zhang. Accelerated replica exchange stochastic gradient langevin diffusion enhanced bayesian deeponet for solving noisy parametric pdes. arXiv preprint arXiv:2111.02484 , 2021

    2021 arXiv

  15. [23]

    Multimodal motion prediction with stacked transformers

    Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7577–7586, 2021

    2021

  16. [24]

    Prose-fd: A mul- timodal pde foundation model for learning multiple operators for forecasting fluid dynamics

    Yuxuan Liu, Jingmin Sun, Xinjie He, Griffin Pinney, Zecheng Zhang, and Hayden Schaeffer. Prose-fd: A mul- timodal pde foundation model for learning multiple operators for forecasting fluid dynamics. arXiv preprint arXiv:2409.09811, 2024

    2024 arXiv

  17. [25]

    Prose: Predicting multiple operators and symbolic expres- sions using multimodal transformers

    Yuxuan Liu, Zecheng Zhang, and Hayden Schaeffer. Prose: Predicting multiple operators and symbolic expres- sions using multimodal transformers. Neural Networks, 180:106707, 2024

    2024

  18. [26]

    Lorsung, Z

    C. Lorsung, Z. Li, and A. B. Farimani. Physics informed token transformer for solving partial differential equations. Machine Learning: Science and Technology , 5(1):015032, 2024

    2024

  19. [27]

    Vilbert: Pretraining task-agnostic visiolinguistic repre- sentations for vision-and-language tasks

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic repre- sentations for vision-and-language tasks. Advances in neural information processing systems , 32, 2019

    2019

  20. [28]

    Deeponet: Learning nonlinear operators for identifying dif- ferential equations based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying dif- ferential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193, 2019

    1910 arXiv

  21. [29]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence , 3(3):218– 229, 2021

    2021

  22. [30]

    Analyzing uncertainty in neural machine translation

    Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. Analyzing uncertainty in neural machine translation. In International Conference on Machine Learning , pages 3956–3965. PMLR, 2018

    2018

  23. [31]

    A survey on transfer learning

    Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2009

    2009

  24. [32]

    A survey on artificial neural networks application for identification and control in environmental engineering: Biological and chemical systems with uncertain models

    Alexander Poznyak, Isaac Chairez, and Tatyana Poznyak. A survey on artificial neural networks application for identification and control in environmental engineering: Biological and chemical systems with uncertain models. Annual Reviews in Control , 48:250–272, 2019

    2019

  25. [33]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  26. [34]

    Ramesh, M

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever. Zero-shot text-to-image generation. In International conference on machine learning , pages 8821–8831. Pmlr, 2021

    2021

  27. [35]

    Schaeffer

    H. Schaeffer. Learning partial differential equations via data discovery and sparse optimization. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , 473(2197):20160446, 2017

    2017

  28. [36]

    Financial time series forecasting with deep learning: A systematic literature review: 2005–2019

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019. Applied soft computing , 90:106181, 2020. 15

    2005

  29. [37]

    Videobert: A joint model for video and language representation learning

    Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464–7473, 2019

    2019

  30. [38]

    J. Sun, Z. Zhang, and H. Schaeffer. Lemon: Learning to learn multi-operator networks. arXiv preprint arXiv:2408.16168, 2024

    2024 arXiv

  31. [39]

    Towards a foundation model for partial differential equation: Multi-operator learning and extrapolation

    Jingmin Sun, Yuxuan Liu, Zecheng Zhang, and Hayden Schaeffer. Towards a foundation model for partial differential equation: Multi-operator learning and extrapolation. arXiv preprint arXiv:2404.12355 , 2024

    2024 arXiv

  32. [40]

    Lxmert: Learning cross-modality encoder representations from transformers

    Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019

    1908 arXiv

  33. [41]

    Are language models actually useful for time series forecasting? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    Mingtian Tan, Mike A Merrill, Vinayak Gupta, Tim Althoff, and Thomas Hartvigsen. Are language models actually useful for time series forecasting? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  34. [42]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M. Lachaux, T. Lacroix, B. Rozi` ere, N. Goyal, E. Hambro, F. Azhar, and others. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  35. [43]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017

    2017

  36. [44]

    H. Wang, Y. Cao, Z. Huang, Y. Liu, P. Hu, X. Luo, Z. Song, W. Zhao, J. Liu, J. Sun, and others. Recent advances on machine learning for computational fluid dynamics: A survey. arXiv preprint arXiv:2408.12171 , 2024

    2024 arXiv

  37. [45]

    Multimodal learning with transformers: A survey

    Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transac- tions on Pattern Analysis and Machine Intelligence , 2023

    2023

  38. [46]

    In-context operator learning with data prompts for differential equation problems

    Liu Yang, Siting Liu, Tingwei Meng, and Stanley J Osher. In-context operator learning with data prompts for differential equation problems. Proceedings of the National Academy of Sciences , 120(39):e2310142120, 2023

    2023

  39. [47]

    Fine-tune language models as multi-modal differential equation solvers

    Liu Yang, Siting Liu, and Stanley J Osher. Fine-tune language models as multi-modal differential equation solvers. arXiv preprint arXiv:2308.05061 , 2023

    2023 arXiv

  40. [48]

    Pde generalization of in-context operator networks: A study on 1d scalar nonlinear conservation laws

    Liu Yang and Stanley J Osher. Pde generalization of in-context operator networks: A study on 1d scalar nonlinear conservation laws. arXiv preprint arXiv:2401.07364 , 2024

    2024 arXiv

  41. [49]

    Z. Ye, X. Huang, L. Chen, Z. Liu, B. Wu, H. Liu, Z. Wang, and B. Dong. Pdeformer-1: A foundation model for one-dimensional partial differential equations. arXiv preprint arXiv:2407.06664 , 2024

    2024 arXiv

  42. [50]

    Scientific large language models: A survey on biological & chemical domains

    Qiang Zhang, Keyang Ding, Tianwen Lyv, Xinda Wang, Qingyu Yin, Yiwen Zhang, Jing Yu, Yuhao Wang, Xiaotong Li, Zhuoyi Xiang, et al. Scientific large language models: A survey on biological & chemical domains. arXiv preprint arXiv:2401.14656 , 2024

    2024 arXiv

  43. [51]

    Bayesian deep operator learning for homogenized to fine-scale maps for multiscale pde

    Zecheng Zhang, Christian Moya, Wing Tat Leung, Guang Lin, and Hayden Schaeffer. Bayesian deep operator learning for homogenized to fine-scale maps for multiscale pde. Multiscale Modeling & Simulation, 22(3):956–972, 2024. A Dataset Details We list below the parametric equation...

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.